We will use Grok 3.5 (maybe we should call it 4), which has advanced reasoning, to rewrite the entire corpus of human knowledge, adding missing information and deleting errors.
Then retrain on that.
Far too much garbage in any foundation model trained on uncorrected data.
This very much depends on the subject, I suspect. For math or computer science, wikipedia is an excellent source, and the credentials of the editors maintaining those areas are formidable (to say the least). Their explanations of the underlaying mechanisms are in my experience a little variable in quality, but I haven’t found one that’s even close to outright wrong.